Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Use a DFS to compare CTFE snapshots #66946

Closed

Conversation

ecstatic-morse
Copy link
Contributor

@ecstatic-morse ecstatic-morse commented Dec 2, 2019

This is an alternative fix to #52475 that hopes to address some parts of #54384.

The major development in this PR is the use of an actual depth-first search to compare the heap between snapshots, instead of the current recursive method. We treat the set of pointers on the stack as an ordered list of "roots". Using this as the initial search stack, a DFS can enumerate all reachable memory in a predictable order. This is important, since without a stable ordering, we would need to compute an isomorphism for the allocation graph of each snapshot to compare them.

Equality between snapshots is checked by iterating in lock-step over their allocation graphs and comparing each Allocation using an IgnoreAllocId wrapper with custom PartialEq and Hash impls. If the allocations are equivalent but for AllocIds, all pointers in the allocation are mapped to a DFS index and then compared.

I find the existing implementation pretty inscrutable, so I'm not equipped to speculate on performance differences. At the very least, my method will use less memory when a collision does occur, since we no longer clone Memory in full, just the reachable portions.

Some aspects of #54384 remain unaddressed, most notably what to do when we fail to resolve an AllocId. My PR takes what I believe is the same approach as the existing implementation; it converts the InterpError into a None and continues traversing the allocation graph. I don't know what the correct behavior would be.

r? @RalfJung
cc @oli-obk

@rust-highfive rust-highfive added the S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. label Dec 2, 2019
@ecstatic-morse ecstatic-morse force-pushed the better-infinite-loop branch 2 times, most recently from 56a760c to 077a810 Compare December 2, 2019 04:44
@ecstatic-morse ecstatic-morse changed the title Use a DFS to compare CTFE snapshots [WIP] Use a DFS to compare CTFE snapshots Dec 2, 2019
/// Erase any allocation IDs.
///
/// Used when comparing heap snapshots.
pub fn erase_id(self) -> Pointer<Tag, ()> {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please rename to erase_alloc_id.

But isn't it easier to just say ptr1.offset == ptr2.offset? Pointers with AllocId type () are kind of pointless (all puns intended).

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Oh, you are recursively doing through even for Place. But how can that make any sense, given that all AllocIds are considered equal at this point...?

Copy link
Contributor Author

@ecstatic-morse ecstatic-morse Dec 2, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

👍 to erase_alloc_id.

This is the part that makes AllocIds compare as equal. It means we can use the derived PartialEq impl. The alternative is to overload AllocIdInvariant for more types, which would also be fine.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

But why does it make any sense to compare just the offsets?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Other code is responsible for comparing the allocation that each AllocId references, specifically the PartialEq implementation of InterpRef. Am I misunderstanding the question?

Comment on lines 471 to 526
if ecx.machine.loop_detector.is_empty() {
// FIXME(#49980): make this warning a lint
ecx.tcx.sess.span_warn(span,
"Constant evaluating a complex constant, this might take some time");
}
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why did you move this out? Seems like a layering violation for this code to care about "emptiness" of the loop detector.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Oh I see, you could remove the tcx and span arguments then. That is a good argument, but is_empty is not a great name.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

has_been_invoked?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No that's negated... is_unused?

Copy link
Contributor Author

@ecstatic-morse ecstatic-morse Dec 2, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

!has_been_invoked is negated but is_unused isn't?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is_being_invoked_the_first_time() 🙃

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

!has_been_invoked is negated but is_unused isn't?

Well I wanted to avoid the ! is all. It's way too easy to miss.

type Item = FrameSnapshot<'a, 'tcx>;
/// A wrapper type for the various interpreter data structures that compares them without
/// considering the numeric value of `AllocId`s.
struct AllocIdInvariant<'a, T>(&'a T);
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

"invariant" is an odd term here, that indicates that something doesn't change.

IgnoreAllocId sounds more accurate? But I have no idea why it even makes any sense to do this.^^

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, I would not be opposed to reverting to the AllocId-sensitive method, or removing this entirely. It is pretty cool when you accidentally omit a break and you're told there's a provably infinite loop in your program, and I have some sentimental connection to this code since it was my first real rust PR. However, I think a configurable instruction limit (with a low default) coupled with a real lint would have basically the same effect.

Once loops are implemented, people will actually start hitting bugs/performance issues in this code. I wanted to make sure I understood what's going on now. I'm worried people will see long-running constant evaluations grind to a halt because of this machinery and have no way to turn it off.

cc @oli-obk

@RalfJung
Copy link
Member

RalfJung commented Dec 2, 2019

Some aspects of #54384 remain unaddressed, most notably what to do when we fail to resolve an AllocId. My PR takes what I believe is the same approach as the existing implementation; it converts the InterpError into a None and continues traversing the allocation graph. I don't know what the correct behavior would be.

This means we have a dangling pointer, right?

let id = stack.pop()?;
let alloc = memory.get_raw(id);

if let Ok(alloc) = alloc {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this where you are ignoring InterpError? This seems reasonable.

We should make sure that we don't catch errors requiring allocation though, similar to this. Maybe create a new helper method at the error type for this?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not sure what you mean by "errors requiring allocation". Are there certain errors that we should bug on instead of ignoring? Is Ub one of these?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Any error type that has a String or other heap types in their fields will allocate when being thrown, and the catching will then destroy the allocated values. If this allocate->throw->catch->destroy happens in a high frequency we're wasting a lot of cycles with all the allocating/deallocating.

So, the idea is to make sure that any error that is caught and acted upon (and not just rethrown) does not allocate

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Idea: Should we make the set of errors that Memory can emit a separate type (with appropriate Into impls) so we know the set of errors that Memory methods can emit and can make sure all of them don't allocate?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That makes sense. I'll probably wait for someone else to do this? It feels like y'all have a decent idea of what you want here.

// dependencies.
let bytes = self.inspect_with_undef_and_ptr_outside_interpreter(all_bytes);
live_allocs.map(|(_, alloc)| alloc.ok().map(AllocIdInvariant))
.eq(other_live_allocs.map(|(_, alloc)| alloc.ok().map(AllocIdInvariant)))
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@RalfJung. I discard the contents of the InterpError with ok. Do you think we can differentiate between null pointers, pointers to ZSTs and dangling pointers by looking at InterpError like you mentioned above? Do ZST pointers have zero-sized backing allocations?

Copy link
Contributor

@oli-obk oli-obk left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I started a review, and forgot about it. I'll continue it later. For now just dumping the little status it has

Comment on lines 471 to 526
if ecx.machine.loop_detector.is_empty() {
// FIXME(#49980): make this warning a lint
ecx.tcx.sess.span_warn(span,
"Constant evaluating a complex constant, this might take some time");
}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is_being_invoked_the_first_time() 🙃

src/librustc_mir/interpret/eval_context.rs Outdated Show resolved Hide resolved
@ecstatic-morse ecstatic-morse force-pushed the better-infinite-loop branch 4 times, most recently from 4027611 to baac4af Compare December 8, 2019 19:41
@ecstatic-morse
Copy link
Contributor Author

@oli-obk, @RalfJung the necessary changes have been made to handle memory cycles. I also added a (very wordy) high-level explanation of the new code in the module docs. Finally, I resolved many of the concerns, although I would still like to do something to address this one.

@ecstatic-morse ecstatic-morse changed the title [WIP] Use a DFS to compare CTFE snapshots Use a DFS to compare CTFE snapshots Dec 8, 2019
src/librustc_mir/interpret/snapshot.rs Outdated Show resolved Hide resolved
//!
//! ```rust,ignore(const-loop)
//! const fn inf_loop() {
//! // Function call to prevent promotion.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

inside const fns promotion is not prevented for const fns, also memoization will memoize this 🙃

You'll want const fn zeros(_: &i32) -> [isize; 4] { [0; 4] } and pass a local variable as a reference at the call site.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Doh! Can't believe I missed this. My worry is that the test for issue 52475 isn't really creating new AllocIds, since the &0 is getting promoted out.

src/librustc_mir/interpret/snapshot.rs Show resolved Hide resolved
ecstatic-morse and others added 2 commits December 9, 2019 13:01
Co-Authored-By: Oliver Scherer <github35764891676564198441@oli-obk.de>
@bors
Copy link
Contributor

bors commented Dec 15, 2019

☔ The latest upstream changes (presumably #67216) made this pull request unmergeable. Please resolve the merge conflicts.

@RalfJung
Copy link
Member

What is the long-term plan here? From other conversation, it seems like one plan was to remove the loop checker entirely. So we probably should finish that discussion before spending more time on this PR?

Also I got quite a backlog of PRs to handle so @oli-obk if you could take over review that would be much appreciated.

Comment on lines +50 to +51
//! Instead of comparing `AllocId`s between snapshots, we first map `AllocId`s in each snapshot to
//! their DFS index, then compare those instead.
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That makes sense. However, what I would have expected then is to see Pointer<Tag, DfsIndex> instead of Pointer<Tag, AllocId>, and then one would compare those. What I still don't get is why we have a way to compare pointers ignoring the AllocId. That means we must have a magic scheme somewhere to figure out all the AllocId we ignored and compare their DfsIndex instead? That sounds quite round-about and fragile, on first sight. (I didn't read the code, though.)

@oli-obk
Copy link
Contributor

oli-obk commented Dec 21, 2019

r? @oli-obk

@rust-highfive rust-highfive assigned oli-obk and unassigned RalfJung Dec 21, 2019
@oli-obk
Copy link
Contributor

oli-obk commented Dec 21, 2019

From other conversation, it seems like one plan was to remove the loop checker entirely. So we probably should finish that discussion before spending more time on this PR?

This was discussed in https://rust-lang.zulipchat.com/#narrow/stream/146212-t-compiler.2Fconst-eval/topic/Time.20limits.20during.20const-eval/near/183588492

and from the last comment

I'm gonna open a PR removing the loop detector after #67260 gets merged.

I am assuming that we can close this PR

@oli-obk oli-obk closed this Dec 21, 2019
Centril added a commit to Centril/rust that referenced this pull request Mar 23, 2020
…op-detector, r=RalfJung

Remove const eval loop detector

Now that there is a configurable instruction limit for CTFE (see rust-lang#67260), we can replace the loop detector with something much simpler. See rust-lang#66946 for more discussion about this. Although the instruction limit is nightly-only, the only practical way to reach the default limit uses nightly-only features as well (although CTFE will still execute code using such features inside an array initializer on stable).

This will at the very least require a crater run, since it will result in an error wherever the "long running const eval" warning appeared before. We may need to increase the default for `const_eval_limit` to work around this.

Resolves rust-lang#54384 cc rust-lang#49980
r? @oli-obk cc @RalfJung
Centril added a commit to Centril/rust that referenced this pull request Mar 23, 2020
…op-detector, r=RalfJung

Remove const eval loop detector

Now that there is a configurable instruction limit for CTFE (see rust-lang#67260), we can replace the loop detector with something much simpler. See rust-lang#66946 for more discussion about this. Although the instruction limit is nightly-only, the only practical way to reach the default limit uses nightly-only features as well (although CTFE will still execute code using such features inside an array initializer on stable).

This will at the very least require a crater run, since it will result in an error wherever the "long running const eval" warning appeared before. We may need to increase the default for `const_eval_limit` to work around this.

Resolves rust-lang#54384 cc rust-lang#49980
r? @oli-obk cc @RalfJung
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
S-waiting-on-review Status: Awaiting review from the assignee but also interested parties.
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet

5 participants